Your client can now produce a decent version of your deliverable without you. Not as good, probably, but competent, structured and available on a Tuesday evening for the price of a subscription. Whatever they are still paying for, it is not the document.
Worth working out what it is, because the answer decides which parts of your firm are load-bearing and which are decoration.
Ronald Coase asked in 1937 why firms exist at all, when markets could in principle coordinate everything. His answer was transaction costs: organising work inside a hierarchy is cheaper than negotiating every task across an open market. The firm exists because internal coordination beats market coordination.
Which makes this moment more interesting than it appears. Agentic systems that draft, route, schedule and execute are driving the cost of explicit coordination towards zero. If coordination were the whole of what a firm provided, the people arguing that firms should flatten and thin would be reading Coase correctly.
But coordination was never what clients were buying. They were buying somebody's willingness to put their name on an answer and be wrong in public if it turns out wrong. That is not a service a machine can render, because the machine has nothing at stake. An approval click from a person who skimmed the output provides the appearance of accountability and none of the substance, and clients can tell the difference sooner than most firms expect.
The firm survives, then, not as a coordination mechanism, which AI replicates, but as a trust mechanism, which it cannot.
Michael Polanyi explained why the judgment underneath cannot be extracted and shipped. He called it tacit knowledge: you can read every book written about riding a bicycle and still fall off, because the knowing is in the riding. The lawyer who reads a contract and senses immediately where the risk sits without being able to say why. The consultant five years into a relationship who knows things about a client's culture that appear in no briefing document. The manager three years into a team who knows exactly who will hold in a crisis. All of it real, all of it valuable, and none of it in the model, because none of it was ever written down.
Here is where the argument turns uncomfortable, because trust is not an abstraction a firm possesses. It is manufactured, continuously, by identifiable people, and those people are the ones currently being described as overhead.
Middle managers are the only people with simultaneous access to the strategic intent of leadership and the operational reality of the work. They do not relay between the two. They translate. An abstract priority arriving from above has to become something meaningful for a team with particular clients, systems and constraints, and the account of frontline reality travelling upward has to become something leadership can act on. Nonaka and Takeuchi described organisational knowledge as a cycle in which tacit understanding is shared, made explicit, combined, then absorbed back into practice. Automated systems are excellent at exactly one phase of that cycle, the combining of explicit structured information, and essentially absent from the rest.
So the flattening argument is half right, which is what makes it dangerous. The half it gets right, explicit coordination, is genuinely automatable. The half it misses is the job.
Eliminate the layer and you do not eliminate the work. You leave it undone, or you redistribute it: downward onto frontline staff who lack the position to do it, and upward onto executives who lack the time and who have now also lost the operational visibility the middle used to supply. The firm does not become flatter and faster. It becomes more fragile and less honest with itself.
Something to try this week. Take one middle manager's scorecard and strike out every line AI can now do: the routing, the status collation, the chasing, the reporting upward. Then look at what is left on the page.
If very little remains, that is a gap in the measurement rather than in the manager. Try adding a single line for something the machine cannot touch: a translation that stopped a team heading the wrong way, someone they brought on, a piece of context that existed nowhere else.
One line is a small start, and I would rather suggest a small one honestly. Changing what people are measured on is slow, awkward work, and I have not yet seen an organisation carry it the whole way through, including ones I have worked with.